Independent Verification
A verifier receives the candidate as a claim to disprove, without the hunter narrative or the desired verdict. Agreement between models is not verification.
A boundary, not a vote
The verifier is an evidence boundary. It should not inherit the hunter confidence, the desired severity, or an instruction that the candidate is true. Agreement without independent reconstruction is not verification.
authorized scope
-> surface map
-> applicable specialist lanes
-> CANDIDATE / UNASSESSED packets
-> coordinator root-cause deduplication
-> neutral verifier packet
-> independent refutation/reconstruction
-> remediation review
-> regression/release verificationIndependence controls
What the verifier receives:
- The candidate claim.
- The necessary evidence locations.
- The scope.
- Compensating-control hints.
What is withheld:
- The proposed truth of the claim.
- The desired severity.
- Hidden evaluation labels.
- The persuasive hunter narrative.
The verifier is then required to navigate and reconstruct independently from the cited source slices. Using a different model or provider helps, but role separation and independent evidence are the actual control.
Verifier classifications
| Classification | Meaning |
|---|---|
| VERIFIED | The evidence chain is complete enough to act on. |
| LIKELY_BUT_UNPROVEN | An important signal missing a decisive link. |
| FALSE_POSITIVE | The suspected path is blocked, or the assumptions are wrong. The reason is retained. |
| DUPLICATE_ROOT_CAUSE | Another verified issue already explains the same failure. |
| BLOCKED_BY_ENVIRONMENT | Proof requires legitimate environment evidence that is unavailable. |
Rejected candidates stay in the report. The false-positive rejection rate is one of the six evaluation metrics, and it cannot be measured if rejections are discarded.
What the verifier attacks
For High and Critical candidates the verifier tries to disprove each of these in turn:
- Attacker control.
- Reachability.
- Missing guard assumptions.
- Role preconditions.
- The claimed impact.
- Whether the vulnerable state is actually producible.
- Whether a compensating control already blocks the exploit.
- Whether the defect is a duplicate of another root cause.
A candidate is not promoted merely because a scanner and a model agree. Compensating controls, unreachable states, missing attacker control, and impossible prerequisites are all valid grounds for rejection.
When independence is degraded
Sometimes the same agent has to verify its own candidate. The framework does not forbid this; it requires the limitation to be recorded.
- Start a clean context and record the reduced-independence limitation in the report.
- If a verifier is not independent, do not describe agreement as independent proof.
- Deduplicate candidates before verification where they obviously share a root cause, and let the verifier identify the uncertain duplicates.
- New evidence can send a candidate back to the owning specialist, but it must not bypass verification.